Data sources were all available PEPFAR-supported African-based HIV-focused household surveys designed to estimate national HIV-1 incidence and other key HIV-related measures, specifically the 95–95-95 targets. We pooled data from 15 surveys collected from 2015–2019: the fifth South African National HIV Prevalence, Incidence, Behaviour, and Communication Survey (SABSSM-V), (23) the Nigeria HIV/AIDS Indicator and Impact Survey (NAIIS), (24) and 13 Population-based HIV Impact Assessment surveys (PHIAs). (25) Surveys came from Southern Africa (Eswatini, Lesotho, Malawi, Namibia, South Africa, Zambia, and Zimbabwe), Eastern Africa (urban Ethiopia, Kenya, Rwanda, Tanzania, and Uganda), and West Africa (Cameroon, Côte d’Ivoire and Nigeria). Surveys from Cameroon and urban regions of Ethiopia were excluded from most analyses because they lacked a question about sexual activity with male partners, a key variable of interest.

All protocols were approved by each in-country local ethics committees and other relevant institutional review boards. Adults 18 years and older provided informed consent. Adolescents 15–17 years-old provided assent after a parent or guardian granted permission. (26) For this analysis, we accessed publicly available de-identified data.

Participants in all surveys completed an individual interview. Participants in 12 surveys provided venous blood for rapid HIV testing and measurement of HIV-1 biomarkers. (27) Participants in SABSSM-V provided dried blood spots for HIV-1 biomarkers. (23) Data were restricted to men aged 15–59 years-old who tested HIV-1 sero-negative in the survey or had recent HIV-1 infection (described below).

Recent HIV-1 infection was our primary outcome of interest. Except South Africa, which used laboratory testing only, all samples were tested for HIV-1 using each country’s HIV rapid testing algorithm and confirmed with laboratory testing. (27) Among those with confirmed HIV, further assessments for recent HIV-1 infection were based on a cross sectional recent HIV-1 infection testing algorithm: laboratory-based limiting antigen-avidity EIA (LAg) normalized optical density values ≤ 1.5, HIV RNA (“viral load”) of ≥1,000 copies/mL, and no qualitative detection of commonly prescribed antiretroviral agents. (28) HIV-1 seropositive individuals without recent HIV-1 infection were excluded.

Twenty-eight individual, partnership, household, and epidemiologic characteristics were harmonized across surveys (Figure 1a–1d). Individual characteristics included age, educational level, marital status, number of wives, occurrence and age of sexual debut, prior HIV-1 testing, and circumcision. Sexual partnership characteristics included number of partners in the last year and whether any of the last three partners was non-steady, transactional, or male; had an unknown HIV status; lived outside the participant’s home; and used a condom at the last encounter. Household variables included urbanicity, number of household members and bedrooms, wall material, water source, toilet type, and item ownership. Three epidemiologic variables were calculated based on survey data from all men and women aged 15–59 years in 208 subnational areas and then applied to each observation (Supplement 1). HIV-1 prevalence was defined as the proportion of adults living with HIV-1 (range=0%-31%). HIV-1 viremia among adults living with HIV (ALHIV HIV-1 viremia) was the proportion of ALHIV with HIV RNA >1000 copies/mL (range=0%-100%). Population HIV-1 viremia was the proportion of the full adult population with HIV-1 RNA >1000 copies/mL (range=0–11%). A map of population HIV-1 viremia has been published. (29) Variables were standardized to have a mean of zero and a standard deviation of one prior to model fitting.

Methods closely follow those used to develop a similar risk assessment tool for women. (29) Supplement 2 provides details on statistical analyses, with a summary provided here. Surveys used a two-stage stratified cluster sample design, which was accounted for in all analyses through incorporation of sampling weights, strata, and clusters. For the data from each national survey, sampling weights were constructed to draw inference to all men aged 15–59 susceptible to HIV-1. Survey data were pooled so that sampling weights summed to population totals by country and age group. Thus, the weighted results represent all men aged 15–59 years who were susceptible to HIV-1 in these 13 countries. Analyses were conducted using SAS version 9.4 (SAS Institute, Cary, NC) and R version 4.2.0 (Vienna, Austria).

Incidence rates were estimated using a cross-sectional recent HIV infection testing algorithm (Supplement 2). The algorithm includes the number of recent HIV-1 infections, number of HIV-1 sero-negative individuals, mean duration of recent infection, assay cutoff time, and proportion of false positive recent cases. (4) Incidence rate estimates were computed using the inctools package in R. For incidence rates and bivariate analyses, variances were estimated using Taylor Series Linearization and used for the construction of 95% confidence intervals (CIs). Missing data for each variable was sparse (0%-5.2%), though 17.8% of observations had ≥1 missing value. Missing values were imputed using multiple imputation by chained equations. (30) The outcome, predictors, and sample design features were included in multiple imputation models. (31) From the weighted imputed datasets, we estimated the proportion of men within each level of each predictor, the percentage of men with recent HIV-1 infection, and odds ratios (approximating incidence rate ratios) with corresponding 95% CIs.

To develop a risk assessment tool, weighted logistic Least Absolute Shrinkage and Selection Operator (Lasso) regression models were applied to the pooled dataset. To avoid over-fitting, five-fold cross validation was used to identify the tuning parameter that maximized the AUC within a 2% tolerance. (32) From the final Lasso model, an overall AUC was estimated by applying the model to the full dataset. The relative importance of the final predictors was estimated by the absolute value of the Lasso model’s coefficients. An additional five-fold cross validation was used to internally validate the model’s AUC. Predictive performance was gauged by the AUC.

Based on the Lasso model, at four sensitivity thresholds (10%, 30%, 70% and 90%) the following quantities were estimated: specificity, number of men prioritized for PrEP, HIV-1 incidence rate among men prioritized for PrEP, and number of men needed to treat (NNT) with PrEP to avert one infection. Annual infections averted were estimated, assuming perfect one-year PrEP effectiveness, which was an upper bound. For all performance metrics, corresponding 95% CIs were constructed based on the Rao-Wu rescaling bootstrap. (33) We also did not account for additional infections averted due to onward transmissions.

Two sensitivity analyses were conducted. First, logistic regression using a backward elimination approach was applied. Based on the estimated bivariate associations between each variable and recent HIV-1 infection, variables with Wald-test p-values <0.05 were included in the initial backward regression model, and then removed individually based on Wald tests, conditional on inclusion of other predictors, until all remaining predictors had p-values ≤0.05. Second, a weighted random forest model was fit to the data. Parameters were tuned to maximize the out of bag five-fold cross validated weighted AUC. For both methods, overall and cross-validated AUCs were estimated.

The unweighted 13-country analytic sample comprised 167,121 men aged 15–59 and represented a target population of all 122 million susceptible men in this age range across these 13 countries between 2015 and 2019. Approximately 22% of the target population came from Southern Africa, 31% from Eastern Africa, and 47% from West Africa (Table 1). In the sample, 112 (0.1%) had recent HIV-1, representing an estimated 256,000 (CI: 160,000–353,000) new annual infections in the target population. The HIV-1 incidence rate was 0.21 per 100 PYs (95% CI: 0.13–0.30) overall: 0.57 per 100 PYs in Southern Africa (95% CI: 0.24–0.91), 0.21 per 100 PYs in Eastern Africa (95% CIs: 0.12–0.30), and 0.05 per 100 PYs in West Africa (95% CI: 0.02–0.08).

In bivariate analysis, at an alpha level of 0.05, two individual characteristics were associated with recent HIV-1 infection: occurrence of sexual debut and prior HIV testing (Figure 1). Partner characteristics associated with recent HIV-1 were number of sexual partners, having a partner living outside the home, condom use, and reporting a male sexual partner. Four household characteristics were associated with recent HIV-1: wall material, number of household members, water source, and vehicle ownership. All three HIV-1 epidemiologic variables were associated with recent HIV-1. HIV-1 prevalence and population HIV-1 viremia were highly correlated and associated with recent HIV-1 in a dose-response relationship.

Only two of 28 variables were retained in the Lasso model: living in an area with high subnational prevalence of HIV-1 viremia and having a male sexual partner. Of the two variables, reporting a male sexual partner had slightly greater importance. AUC was 0.80 (0.71–0.89) (Figure 2) and cross-validated AUC was 0.76 (0.64–0.87). When applied to each subregion, overall AUC was 0.64 in Southern Africa, 0.76 in Eastern Africa, and 0.63 in West Africa.

Based on the full-population Lasso model, we examined tradeoffs between reach (cases averted) and efficiency (NNT) at four sensitivity levels along the ROC curve from 10% to 90% (Table 2). Preventing more infections required more men on PrEP and worse efficiency. At a sensitivity of 10%, 357,000 men would be prioritized for PrEP with an HIV-1 incidence of 7.0 per 100 PYs (95% CI: 0.0–15.7). If these men were perfectly adherent, up to 25,000 cases could be averted each year, and NNT would be 14. At the other extreme, a sensitivity of 90%, 50.2 million men (41% of all susceptible adult men) would be prioritized for PrEP; if these men were perfectly adherent, most new HIV-1 infections (N=229,000 cases) could be averted, but NNT would be 219.

We mapped which men would be prioritized for PrEP at each sensitivity threshold (Figure 3). At a sensitivity of 10%, PrEP would be prioritized exclusively for men who have sex with men (MSM): nearly all MSM in Southern Africa, many MSM in Eastern Africa, and few MSM in West Africa. To prevent 90% of new infections, PrEP would be offered to nearly all men aged 15–59 years in Southern Africa, a subset of MSM and other men in Eastern Africa, and all MSM in Western Africa.

The surveys in Cameroon and Ethiopia were excluded from the main analyses because they did not include a question about having a male sexual partner. In a 15-country Lasso model that included 27 of the 28 variables, only population HIV-1 viremia was retained. Both the overall AUC and cross-validated AUC were 0.78, similar to the main model. However, the prioritization map was different. (Supplement 3).

Predictive performance of both backwards regression and the random forest approaches was similar to the Lasso models, but less parsimonious. With backwards regression, six variables were retained, including living in an area with high HIV-1 prevalence and reporting a male sexual partner. Overall AUC was 0.83; cross-validated AUC was 0.74 (Supplement 4). The random forest model retained 26 of 28 candidate predictors and achieved an overall AUC of 0.87 and a cross-validated AUC of 0.77 (Supplement 5). In both models, higher overall AUC relative to cross-validated AUC may be due to overfitting.